Sustainable AI report looks at power, open source hardware

The UK’s National Engineering Policy Centre has highlighted areas where sustainable AI will impact on the development of semiconductors and data centres, from the power consumption to the water used for cooling and the repair and reuse of open source hardware.

This is the first report on the sustainability of AI from the Centre, which brings together the Royal Academy of Engineering, the IET and the Chartered Institute of IT.

The report highlights that AI can be used to help optimise energy consumption, manage grid demand, reduce waste, monitor ecosystem health, and quantify the impact of climate change and adaptation strategies.

However to maximise the environmental and societal benefits that AI can offer, systems and services must be environmentally sustainable on a lifecycle basis. As AI systems and services are designed, built and used, they place demands on resources such as energy, water and critical materials.

The major issue identified is that by driving up energy demand, AI systems and services can increase the challenge of moving to a decarbonised electricity system. There are actions that can be taken now, across the AI value chain, to better understand and reduce this unsustainable resource consumption.

This comes as the UK government supports £14bn of investment in AI over the next ten years.

There is also more demand for data centre power. In 2023 there was 601 megawatts of new demand for data centres across the 14 largest markets in Europe, up from 546 megawatts in 2022. In the same period, 561 megawatts of new supply was delivered. This was the second time in five years that new demand exceeded new supply in Europe. This, in turn, led to data centre vacancy rates in the top five European markets (Frankfurt, London, Amsterdam, Paris, Dublin), hitting an all-time low of 10.6% at the end of 2023

More demand for state-of-the-art hardware: growth in semiconductor manufacturing is anticipated to continue, partially driven by demand for specialised hardware to support AI workloads. For example, generative AI chips are predicted to have reached more than $50 billion in sales for 2024. In the longer term, it is projected that AI chips could reach $400 billion in sales by 2027, although this would account for almost half of the entire semiconductor market at that time, which is unlikely.

This is driven by the increasing size of AI models, which have grown 4.4 times a year since 2010, requiring increasing resources for training. If model training continues to follow current trends, that models could be trained on datasets roughly equal in size to the available stock of public human text data as early as 2026. This is also driving demand for data storage: data centre and endpoint storage capacity is predicted to grow from 10.1 zettabytes in 2023 to 21 zettabytes in 2027, partially driven by AI demand.

There is also a drive to more resource-intensive inference: increases in the size of inputs (represented by token limits) and the number of model parameters are resulting in more resource consumption by users. While a shift towards smaller models may impact this trend, where smaller specialist models are assembled as part of an agent based system, the overall impacts may be comparable to that of existing generative AI models.

All this is leading to innovation in developing alternative models that can drive reductions in data storage and compute requirements while delivering comparable, or even improved, performance on certain tasks.

 Alternative chip architectures, meanwhile, can potentially reduce processing time and power requirements for AI workloads, as highlighted by UK startup Fractile.

Combined, or in isolation, new models and chips may significantly reduce the environmental impacts that AI systems and services produce. Both alternative models and chips are active areas of research, within which several approaches with varying timelines for deployment are being developed.

Appropriate incentives to prioritise environmental awareness are likely to increase the appetite for alternative models and chips. Smaller models: smaller, task-specific models which can be trained from scratch, or developed by tuning or distilling a larger model, use smaller datasets and less compute than larger multipurpose models, often without a significant drop-off in performance.

For example small language models (those not exceeding 7 billion parameters), are currently being deployed on end-user or edge devices. These small language models can deliver comparable, or even superior performance, to large language models.

With all this in mind, the report highlights five areas for sustainable AI:  

Expanding environmental reporting mandates would cover energy consumption, energy sources, water consumption and withdrawal, water sources, carbon emissions, e-waste recycling and the Power Usage Effectiveness (PUE) of data centres and make reporting mandatory. These requirements should be based on Industry-based Guidance on implementing Climate-related Disclosures.

This would need new open and trustworthy data collection and reporting protocols and tools to enable data sharing across the value chain for producers, providers and partners of all sizes. These could be shared through a proposed “AI Knowledge Hub”.

This would also include requirements for reporting on the reuse of server equipment. A 2021 study indicated that 28% of companies globally reuse or repurpose IT hardware internally – and initiatives such as the Circular Electronics Partnership (CEP) actively promote circular practice.

Addressing information asymmetries across the value chain to give users more understanding of the energy consumption and sustainability of their use of AI.

Setting environmental sustainability requirements for datacentres would reduce the use of drinkable water for all data centre activities and eliminate its use for cooling. This should be supported by best-practice guidance in ISO 46001:2019 (Water efficiency management systems) and ‘zero potable water for cooling’ commitments from data centre operators.

Maximising the reuse and recycling of server equipment can use existing standards for repair and refurbishment, as well as by developing targets for hardware reuse through open-source hardware designs.

Reconsidering data collection, transmission, storage, and management practices cites the development of the National Data Library as an opportunity to drive environmentally sustainable data management, says the report. This could include private data, models and model components as this could significantly reduce the environmental impacts of development and storage.

The fifth area identified is government investment through long term procurement deals and a focus on smaller models.

The report was put together after a workshop and a series of interviews with experts from industry, academia, civil society and policymakers. The recommended actions aim to help make the UK a leader in AI sustainability and minimise the risk of lock-in to systems that are not sustainable in the long term.

The centre is working on further actions for environmentally sustainable AI which can build on these steps.

The report is here,